File size: 17,200 Bytes
932bc69
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
#!/usr/bin/env python3
"""Prepare a deterministic ranked Infinity-Parser2 dataset for training."""

from __future__ import annotations

import argparse
import json
import os
import shutil
from collections.abc import Mapping
from pathlib import Path
from typing import Any

from datasets import Dataset, load_from_disk

from speculators.data_generation.preprocessing import (
    build_speculator_training_dataset,
    load_processor,
    load_raw_dataset,
)
from speculators.train.vocab_mapping import save_token_frequency_distribution

SELECTION_NAME = "ranked_selection.jsonl"
MANIFEST_NAME = "ranked_selection.json"
SHA256_HEX_LENGTH = 64
_PRESERVED_COLUMNS = ("_id", "_source_line_index", "_candidate_rank")
_HASH_COLUMNS = (
    "_sample_manifest_sha256",
    "_generation_config_sha256",
)
_OVERWRITE_FILES = {
    "dataset_info.json",
    "state.json",
    "token_freq.pt",
    SELECTION_NAME,
    MANIFEST_NAME,
}


def _require_sha256(value: object, field: str) -> str:
    if (
        not isinstance(value, str)
        or len(value) != SHA256_HEX_LENGTH
        or any(character not in "0123456789abcdef" for character in value)
    ):
        raise ValueError(f"{field} must be a lowercase SHA-256 digest")
    return value


def _flatten_ranked_metadata(row: Mapping[str, Any]) -> dict[str, Any]:
    record_id = row.get("id")
    provenance = row.get("provenance")
    if not isinstance(record_id, str) or not record_id:
        raise ValueError("each ranked row must contain a non-empty string id")
    if not isinstance(provenance, Mapping):
        raise ValueError("each ranked row must contain provenance metadata")

    source_index = provenance.get("source_line_index")
    candidate_rank = provenance.get("candidate_rank")
    if type(source_index) is not int or source_index < 0:
        raise ValueError("source_line_index must be a non-negative integer")
    if type(candidate_rank) is not int or candidate_rank < 0:
        raise ValueError("candidate_rank must be a non-negative integer")

    return {
        "_id": record_id,
        "_source_line_index": source_index,
        "_candidate_rank": candidate_rank,
        "_sample_manifest_sha256": _require_sha256(
            provenance.get("sample_manifest_sha256"),
            "sample_manifest_sha256",
        ),
        "_generation_config_sha256": _require_sha256(
            provenance.get("generation_config_sha256"),
            "generation_config_sha256",
        ),
    }


def _single_value(dataset: Dataset, column: str) -> str:
    values = dataset.unique(column)
    if len(values) != 1:
        raise ValueError(f"{column.removeprefix('_')} is not consistent")
    return str(values[0])


def _validate_ranked_rows(dataset: Dataset) -> tuple[str, str]:
    if len(dataset) == 0:
        raise ValueError("the ranked regeneration pool is empty")

    ranks = dataset.unique("_candidate_rank")
    if len(ranks) != len(dataset):
        raise ValueError("candidate_rank values are not unique")
    record_ids = dataset.unique("_id")
    if len(record_ids) != len(dataset):
        raise ValueError("record ids are not unique")

    return (
        _single_value(dataset, "_sample_manifest_sha256"),
        _single_value(dataset, "_generation_config_sha256"),
    )


def prepare_ranked_dataset(args: argparse.Namespace) -> Dataset:
    """Tokenize the ranked reserve and select the first surviving target rows."""
    if args.ranked_target_samples <= 0:
        raise ValueError("ranked_target_samples must be positive")
    if not 0.0 < args.token_freq_train_ratio <= 1.0:
        raise ValueError("token_freq_train_ratio must be in (0, 1]")
    if args.preprocessing_batch_size <= 0:
        raise ValueError("preprocessing_batch_size must be positive")

    raw_dataset, normalize_fn = load_raw_dataset(str(args.data))
    if normalize_fn is not None:
        raw_dataset = raw_dataset.map(
            normalize_fn,
            num_proc=args.num_preprocessing_workers,
            keep_in_memory=False,
        )
    raw_dataset = raw_dataset.map(
        _flatten_ranked_metadata,
        num_proc=args.num_preprocessing_workers,
        keep_in_memory=False,
        batch_size=args.preprocessing_batch_size,
        desc="Validating ranked regeneration metadata",
    )
    sample_sha256, generation_sha256 = _validate_ranked_rows(raw_dataset)
    raw_dataset = raw_dataset.sort(
        "_candidate_rank",
        keep_in_memory=False,
        writer_batch_size=args.preprocessing_batch_size,
    )

    processor = load_processor(
        args.model,
        trust_remote_code=args.trust_remote_code,
    )
    processed = build_speculator_training_dataset(
        dataset=raw_dataset,
        processor=processor,
        max_length=args.seq_length,
        num_proc=args.num_preprocessing_workers,
        render_endpoint=args.render_endpoint,
        local_render=args.local_render,
        drop_clipped_rows=args.drop_clipped_rows,
        minimum_valid_tokens=args.minimum_valid_tokens,
        preserve_columns=_PRESERVED_COLUMNS,
        keep_in_memory=False,
        map_batch_size=args.preprocessing_batch_size,
        render_chat_template_kwargs={"enable_thinking": False},
    ).with_format(None)

    eligible_ranks = sorted(int(rank) for rank in processed.unique("_candidate_rank"))
    eligible_records = len(eligible_ranks)
    if eligible_records < args.ranked_target_samples:
        raise ValueError(
            "preprocessing survivors are below target: "
            f"{eligible_records} < {args.ranked_target_samples}"
        )

    selected_ranks = eligible_ranks[: args.ranked_target_samples]
    selected_cutoff = selected_ranks[-1]
    selected = processed.filter(
        lambda ranks: [rank <= selected_cutoff for rank in ranks],
        input_columns=["_candidate_rank"],
        batched=True,
        batch_size=args.preprocessing_batch_size,
        num_proc=args.num_preprocessing_workers,
        keep_in_memory=False,
        desc="Selecting ranked records after assistant-turn fan-out",
    )

    train_records = int(args.ranked_target_samples * args.token_freq_train_ratio)
    if train_records <= 0:
        raise ValueError("token_freq_train_ratio leaves an empty training prefix")
    train_cutoff = selected_ranks[train_records - 1]
    token_frequency_rows = selected.filter(
        lambda ranks: [rank <= train_cutoff for rank in ranks],
        input_columns=["_candidate_rank"],
        batched=True,
        batch_size=args.preprocessing_batch_size,
        num_proc=args.num_preprocessing_workers,
        keep_in_memory=False,
        desc="Selecting record prefix for token frequencies",
    )

    selected = selected.rename_columns(
        {
            "_id": "id",
            "_source_line_index": "source_line_index",
            "_candidate_rank": "candidate_rank",
        }
    )
    selected.info.description = json.dumps(
        {
            "ranked_preprocessing": {
                "dataset_order": "candidate_rank_ascending_assistant_turn",
                "eligible_records": eligible_records,
                "eligible_training_rows": len(processed),
                "generation_config_sha256": generation_sha256,
                "sample_manifest_sha256": sample_sha256,
                "selected_records": args.ranked_target_samples,
                "selected_training_rows": len(selected),
            }
        },
        sort_keys=True,
    )

    token_frequency_rows.set_format(type="torch")
    save_token_frequency_distribution(
        dataset=token_frequency_rows,
        output_path=Path(args.output) / "token_freq.pt",
    )
    selected.set_format(type="torch")
    return selected


def _ranked_metadata(dataset: Dataset) -> dict[str, Any]:
    try:
        description = json.loads(dataset.info.description or "{}")
        metadata = description["ranked_preprocessing"]
    except (json.JSONDecodeError, KeyError, TypeError) as exc:
        raise ValueError("dataset is missing ranked preprocessing metadata") from exc
    if not isinstance(metadata, dict):
        raise ValueError("ranked preprocessing metadata must be an object")
    return metadata


def _unique_selection_rows(dataset: Dataset) -> list[tuple[int, int, str]]:
    """Collapse contiguous assistant-turn rows back to their source records."""
    plain = dataset.with_format(None)
    rows: list[tuple[int, int, str]] = []
    previous: tuple[int, int, str] | None = None
    for source_index, candidate_rank, record_id in zip(
        plain["source_line_index"],
        plain["candidate_rank"],
        plain["id"],
        strict=True,
    ):
        current = (int(source_index), int(candidate_rank), str(record_id))
        if previous is not None and current[1] == previous[1]:
            if current != previous:
                raise ValueError(
                    "assistant-turn rows disagree on source record metadata"
                )
            continue
        if previous is not None and current[1] < previous[1]:
            raise ValueError("prepared rows are not ordered by candidate rank")
        rows.append(current)
        previous = current
    return rows


def stage_selection(
    dataset: Dataset,
    *,
    output: Path,
    target_samples: int,
) -> tuple[Path, dict[str, Any]]:
    """Write the selection payload to a partial file before dataset publish."""
    rows = _unique_selection_rows(dataset)
    if len(rows) != target_samples:
        raise ValueError(
            f"selected dataset has {len(rows)} records, expected {target_samples}"
        )
    metadata = _ranked_metadata(dataset)
    sample_sha256 = _require_sha256(
        metadata.get("sample_manifest_sha256"),
        "sample_manifest_sha256",
    )
    generation_sha256 = _require_sha256(
        metadata.get("generation_config_sha256"),
        "generation_config_sha256",
    )
    if metadata.get("selected_records") != target_samples:
        raise ValueError("dataset metadata has a different selected record count")
    if metadata.get("selected_training_rows") != len(dataset):
        raise ValueError("dataset metadata has a different training row count")

    output.mkdir(parents=True, exist_ok=True)
    staged = output / f".{SELECTION_NAME}.{os.getpid()}.partial"
    staged.unlink(missing_ok=True)
    rows.sort()
    with staged.open("wt", encoding="utf-8") as handle:
        for source_index, candidate_rank, record_id in rows:
            handle.write(
                json.dumps(
                    {
                        "id": record_id,
                        "source_line_index": source_index,
                        "candidate_rank": candidate_rank,
                    },
                    ensure_ascii=False,
                    sort_keys=True,
                    separators=(",", ":"),
                )
                + "\n"
            )

    manifest = {
        "complete": True,
        "target_records": target_samples,
        "training_rows": len(dataset),
        "sample_manifest_sha256": sample_sha256,
        "generation_config_sha256": generation_sha256,
        "selection": {
            "path": SELECTION_NAME,
            "records": target_samples,
        },
    }
    return staged, manifest


def _write_json_atomic(path: Path, value: Mapping[str, Any]) -> None:
    staged = path.with_name(f".{path.name}.{os.getpid()}.partial")
    staged.write_text(
        json.dumps(value, ensure_ascii=False, sort_keys=True, indent=2) + "\n",
        encoding="utf-8",
    )
    os.replace(staged, path)


def publish_selection(
    output: Path,
    *,
    staged_selection: Path,
    manifest: Mapping[str, Any],
) -> None:
    """Publish the selection and write its completion manifest last."""
    os.replace(staged_selection, output / SELECTION_NAME)
    _write_json_atomic(output / MANIFEST_NAME, manifest)


def validate_existing_output(args: argparse.Namespace) -> None:
    """Validate a previously completed output before treating it as reusable."""
    output = Path(args.output)
    manifest_path = output / MANIFEST_NAME
    if not manifest_path.is_file():
        raise ValueError(f"existing output is incomplete: missing {manifest_path}")
    try:
        manifest = json.loads(manifest_path.read_text(encoding="utf-8"))
    except (OSError, json.JSONDecodeError) as exc:
        raise ValueError(f"cannot read selection manifest: {manifest_path}") from exc
    selection = manifest.get("selection")
    training_rows = manifest.get("training_rows")
    if (
        manifest.get("complete") is not True
        or manifest.get("target_records") != args.ranked_target_samples
        or type(training_rows) is not int
        or training_rows < args.ranked_target_samples
        or not isinstance(selection, dict)
        or selection.get("path") != SELECTION_NAME
        or selection.get("records") != args.ranked_target_samples
    ):
        raise ValueError("existing prepared output was created with different settings")

    selection_path = output / SELECTION_NAME
    if not selection_path.is_file():
        raise ValueError(f"existing output is missing {selection_path}")
    with selection_path.open("rt", encoding="utf-8") as handle:
        selection_records = sum(1 for line in handle if line.strip())
    if selection_records != args.ranked_target_samples:
        raise ValueError("existing selection has a different record count")

    dataset = load_from_disk(output)
    if len(dataset) != training_rows:
        raise ValueError("existing dataset has a different training row count")
    metadata = _ranked_metadata(dataset)
    if (
        metadata.get("selected_records") != args.ranked_target_samples
        or metadata.get("selected_training_rows") != training_rows
    ):
        raise ValueError("existing dataset metadata has different record counts")


def _assert_safe_to_overwrite(output: Path) -> None:
    unexpected = []
    for path in output.iterdir():
        if path.is_file() and (
            path.suffix == ".arrow"
            or path.name in _OVERWRITE_FILES
            or (path.name.startswith(".") and path.name.endswith(".partial"))
        ):
            continue
        unexpected.append(path)
    if unexpected:
        paths = ", ".join(str(path) for path in unexpected)
        raise ValueError(
            "--overwrite would remove files outside the prepared-data "
            f"artifacts: {paths}"
        )


def build_parser() -> argparse.ArgumentParser:
    parser = argparse.ArgumentParser(
        description="Prepare ranked Infinity-Parser2 regeneration data"
    )
    parser.add_argument("--model", required=True)
    parser.add_argument("--data", type=Path, required=True)
    parser.add_argument("--output", type=Path, required=True)
    parser.add_argument("--seq-length", type=int, default=20480)
    parser.add_argument("--ranked-target-samples", type=int, required=True)
    parser.add_argument("--token-freq-train-ratio", type=float, default=0.99)
    render = parser.add_mutually_exclusive_group(required=True)
    render.add_argument("--render-endpoint")
    parser.add_argument(
        "--drop-clipped-rows",
        action="store_true",
        help="drop rows longer than --seq-length instead of truncating them. "
        "Required for online training, where the stored conversation is "
        "re-rendered and a truncated row's ids can never match; offline "
        "and text runs should leave it off and keep truncating.",
    )
    render.add_argument(
        "--local-render",
        action="store_true",
        help="tokenize with the processor in-process instead of calling the "
        "vLLM /render endpoint; same ids, far cheaper for image corpora",
    )
    parser.add_argument("--minimum-valid-tokens", type=int)
    parser.add_argument("--num-preprocessing-workers", type=int, default=8)
    parser.add_argument("--preprocessing-batch-size", type=int, default=1000)
    parser.add_argument("--trust-remote-code", action="store_true")
    parser.add_argument("--overwrite", action="store_true")
    return parser


def main() -> int:
    args = build_parser().parse_args()
    output = args.output.expanduser().resolve()
    args.output = output
    if output.exists():
        if args.overwrite:
            _assert_safe_to_overwrite(output)
            shutil.rmtree(output)
        elif any(output.iterdir()):
            validate_existing_output(args)
            print(f"prepared dataset already complete: {output}")
            return 0
    output.mkdir(parents=True, exist_ok=True)

    dataset = prepare_ranked_dataset(args)
    staged_selection, manifest = stage_selection(
        dataset,
        output=output,
        target_samples=args.ranked_target_samples,
    )
    try:
        dataset.save_to_disk(output)
        publish_selection(
            output,
            staged_selection=staged_selection,
            manifest=manifest,
        )
    except Exception:
        staged_selection.unlink(missing_ok=True)
        raise
    print(
        f"prepared {len(dataset)} assistant-turn rows from "
        f"{args.ranked_target_samples} ranked records: {output}"
    )
    return 0


if __name__ == "__main__":
    raise SystemExit(main())